Tavily

What is Tavily?

Tavily is a web search and retrieval API built for AI agents rather than for people. The distinction matters more than it sounds.

A person searching the web issues one query, scans a page of links, opens two of them and decides. An agent does not work that way. It searches, reads, decides it does not have enough, searches again, and repeats until it can answer. One user question routinely becomes six or eight retrieval calls before a single token of the answer is written.

That changes what a search API has to be good at. Human search optimises for the first page of results, because a person is going to read it. Agent search optimises for what happens after the result comes back, because nothing is going to read it except a model with a token budget and a latency deadline.

Tavily was acquired by Nebius in February 2026.

Tavily at a glance

Tavily main AI capabilities

Most search APIs hand back links. For a person that is the answer. For an agent it is a to-do list: fetch each page, strip the navigation and the cookie banner, handle the paywall, cut what is left to fit the context window, and only then start reasoning. That work is invisible in a demo and it is most of your latency in production.

Tavily returns extracted, ranked content in the same call. The agent gets something it can reason over immediately rather than a set of URLs it has to go and process itself. That removes a stage from every pass of the loop, and the loop is where the cost lives.

Tavily has also published a Company Search Benchmark built on real agent tasks, in which it ranks first among the nine providers it evaluated, according to Tavily. The method is in Tavily's launch post and the dataset is public on Hugging Face.

When should you choose Tavily?

Choose Tavily if:

You are building an agent that searches repeatedly within a single user request, and the retrieval step is already a visible share of your latency or your token bill.

You want one provider for the whole retrieval surface. Search, scraping, site mapping, crawling and deep research are five model strings against the same provider, at the same price, rather than five integrations.

You want a synthesised answer rather than raw results. Deep research returns a written answer with its sources, which removes the orchestration layer most teams build themselves.

Consider another provider if:

Your workload is mostly crawls or deep research at high volume. These capabilities use several credits per job, so estimate the cost per job on a sample before scaling.

You need capabilities Tavily does not cover on Eden AI, such as batch scraping or schema-driven structured extraction.

You need EU data residency today. Tavily offers EU query processing through a dedicated endpoint, with servers located in Germany and France, and access is currently limited to selected customer evaluations. On Eden AI, Tavily runs in the US region for now, and an EU endpoint is coming soon.

Tavily pros and cons

Tavily features and capabilities on Eden AI

Available Tavily capabilities

Sync vs async Tavily endpoints on Eden AI

Search, scraping and site map return in the request. Crawl and deep research are long-running, so they start a job and you poll for the result. Plan for that in the agent loop: an async call is not a drop-in replacement for a synchronous one, and a research step that takes a minute needs somewhere to wait that is not the user's request thread.

How Tavily is priced on Eden AI

Tavily is billed on Eden AI at $0.008 per Tavily credit. A basic search uses one credit, so it costs $0.008. Site map, crawl and deep research use more credits depending on how much work they do. In our tests on 22 October 2026, a site map cost $0.016, a three-page crawl $0.048 and a deep research answer $0.112.

What can you build with Tavily?

Research agents that decide when they have enough

The loop is the point. An agent that can search, read, judge the result insufficient and search again produces a different quality of answer from one that runs a single query. Deep research packages that loop server-side, which is the fastest way to find out whether the pattern is worth building yourself.

RAG pipelines with live context

A knowledge base goes stale the day it is built. Adding a retrieval step for anything time-sensitive, such as prices, regulation or product availability, keeps the answer current without reindexing anything.

Monitoring and enrichment workflows

Site map and crawl handle the case where you know the site but not the pages. Map a domain, crawl what matters, and feed the content to a model for extraction, all through the same key as the model itself.

Why use Tavily through Eden AI?

Retrieval sits beside the models that consume it rather than in a separate integration with its own key, its own bill and its own vendor review.

The practical consequence is measurement. Finding out whether your agent performs better on one retrieval provider or another means running the same queries through both, which normally costs an account and a contract per provider. That is why almost nobody does it properly. Through Eden AI it is a string change:

curl -X POST https://api.edenai.run/v3/universal-ai \
  -H "Authorization: Bearer $EDENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "web/search/tavily",
    "input": {
      "query": "EU AI Act obligations for general purpose model providers",
      "max_results": 5
    }
  }'

Eden AI also handles fallbacks, so a provider having a bad afternoon becomes a retry rather than a failed request. For a retrieval layer called several times per user question, that matters more than any benchmark.

Frequently asked questions about Tavily on Eden AI